inpcb: return ENOMEM if bind(2) fails to allocate lbgroup

[glebius]

Summary

The SO_REUSEPORT_LB isn't a standard option, neither ENOMEM is a specified
return code from bind(2), but it definitely is more appropriate than
EAGAIN or the masked ENOBUFS.

[markj]

sys/netinet/in_pcb.c

736

What's the purpose of this assertion? If you want to document that in_pcbinshash() can only fail for lbgroup sockets, IMO it would be better to somehow document that fact in in_pcbinshash().

[glebius]

sys/netinet/in_pcb.c

736

Yes, that's the purpose of course. How do want this be documented in in_pcbinshash()? Like in the comment above the function?

P.S. We potentially can pre-allocate memory at the beginning of the bind(2) syscall and make it never fail even with SO_REUSEPORT_LB. Or pre-allocate it at setsockopt(2) stage. But I don't think the added complexity is worthy. If you are running a load balanced web server and you end up with kernel memory so low that such a tiny malloc(9) fails, you already are screwed up too bad. So bad, that non-standard ENOMEM from bind(2) no longer is an issue :)

[markj]

sys/netinet/in_pcb.c

736

My point about the assertion is that it's a bit fragile. If someone changes in_pcbinshash() such that the assertion is wrong, they might not notice it. I'd just add it as a comment above in_pcbinshash().

IMO it's ugly for a system call to fail because of a transient memory shortage, but indeed, I'm not sure the complexity of fixing it is worth it. I do consider it a bug, though.

[glebius]

sys/netinet/in_pcb.c

736

Got it. I will add comment above in_pcbinshash().

Note that this revision depends on D49151 to proceed.